Cu-ag alloy performance prediction method combining multi-pass deformation and recovery evolution characteristics

Abstract

The application discloses a Cu-Ag alloy performance prediction method combining multi-pass deformation and recovery evolution characteristics, and the method comprises the following steps: constructing an original data set and completing standardization purification treatment; relying on thermodynamics and dynamics mechanism to quantize microstructure evolution characteristics corresponding to multi-pass deformation and annealing recovery; constructing a machine learning prediction model; completing model accuracy and interpretability verification through multi-dimensional indexes; inputting Cu-Ag alloy composition ratio, each pass cold deformation, annealing temperature and holding time of the Cu-Ag alloy to be measured into the machine learning prediction model, and outputting electrical conductivity, hardness, yield strength, tensile strength and fracture elongation rate. The application solves the technical problems that the existing Cu-Ag alloy performance prediction scheme is dependent on a pure data "black box" model, does not combine material microstructure evolution mechanism, has poor interpretability, generalization ability is limited under complex multi-pass process, prediction accuracy is insufficient, and traditional experimental trial-and-error development cycle is long, cost is high and the like.

Description

Technical Field

[0001] This invention relates to the field of alloy performance testing technology, specifically to a method for predicting the performance of Cu-Ag alloys by combining multi-pass deformation and recovery evolution characteristics. Background Technology

[0002] Cu-Ag alloys belong to a typical binary eutectic system and are currently the only copper alloy system that can achieve ultra-high strength (1000MPa-1500MPa) through physical processing while maintaining excellent electrical conductivity (65%-80% IACS). Because they can simultaneously maintain extremely high tensile strength and excellent electrical conductivity after large deformation processing, they have broad application prospects in fields such as power engineering, strong magnetic field devices, aerospace, and high-power electronic devices.

[0003] For high-performance Cu-Ag alloys, traditional experimental trial-and-error methods remain the mainstream performance evaluation approach. This method heavily relies on engineering experience and large-scale physical experiments, exhibiting drawbacks such as low efficiency, high cost, and insufficient process parameter optimization capabilities during the research and development process. It struggles to meet the diverse, customized, and rapidly evolving development needs of contemporary alloy materials. Particularly when dealing with alloy systems coupled with multiple compositions and complex processes, the relationship between process, microstructure, and properties exhibits highly nonlinear characteristics, making it difficult for traditional experience-based methods to achieve efficient and accurate process optimization within a broad parameter space.

[0004] With the rapid development of materials genome engineering, computational materials science, and machine learning technologies, novel methods for predicting alloy properties are constantly emerging. Currently, there are abundant patent achievements in the field of alloy property prediction and optimization, but specialized research on Cu-Ag binary alloy systems is still relatively scarce. Existing prediction methods mostly rely on "black box" models, which, due to the lack of in-depth characterization of the microscopic evolution of materials, result in insufficient interpretability and limited generalization ability under complex processing systems. Summary of the Invention

[0005] This invention provides a Cu-Ag alloy performance prediction method that combines multi-stage deformation and recovery evolution characteristics. The aim is to solve the technical problems of existing prediction methods, which mostly rely on "black box" models, lack in-depth characterization of the microscopic evolution law of materials, have insufficient model interpretability, and have limited generalization ability under complex process systems. This invention aims to achieve high-precision characterization of performance by introducing mechanistic features in the processing and manufacturing process, and significantly improve the iterative efficiency of material design and process optimization.

[0006] The technical solution provided by this invention is as follows: This invention provides a method for predicting the properties of Cu-Ag alloys by combining the characteristics of multi-stage deformation and recovery evolution, comprising the following steps: A raw dataset covering the entire process of Cu-Ag alloy composition, process, and performance is constructed. Outlier removal and standardization preprocessing are performed on the raw dataset to obtain the first dataset. Based on the heat treatment and multi-pass deformation and recovery softening feature data collected during the construction of the original dataset, calculations are performed to quantify and generate the mechanistic features of multiple types of microstructure evolution, resulting in a second dataset. Based on the first dataset and the second dataset, machine learning prediction models are constructed and trained for the target performance metrics, respectively. The performance of the machine learning prediction model was evaluated by combining cross-validation and independent test sets, and the rationality and interpretability of the model mechanism were verified by feature analysis. By inputting the Cu-Ag alloy composition ratio, cold deformation amount of each pass, annealing temperature and holding time of the Cu-Ag alloy to be tested into the machine learning prediction model, the electrical conductivity, hardness, yield strength, tensile strength and elongation at break are output.

[0007] Optionally, the calculation based on the heat treatment and multi-pass deformation and recovery softening feature data collected during the construction of the original dataset quantifies and generates mechanistic features of multiple types of microstructure evolution, including: Initial solid solution potential energy characteristics were extracted. By combining the single-pass annealing temperature, holding time, and recovery activation energy, the thermo-softening factor of each aging pass was extracted. The strain-recovery alternating recursive equation is used to correct the residual strain of the previous pass and superimpose the intervention strain of the current pass. Nonlinear square root mapping of effective cumulative strain across multiple channels yields dislocation strengthening characteristics.

[0008] Optionally, the method for extracting the initial solid solution potential energy characteristics is as follows: The initial solid solution potential energy characteristics are extracted using the Arrhenius equation, which couples thermodynamics and kinetics. The expression of the Arrhenius equation is as follows:

[0009] In the formula, Let be the solid solution activation energy of the alloy, R be the ideal gas constant, k be the diffusion kinetic rate constant, and A1 be the material comprehensive pre-factor. The absolute temperature for solution / homogenization. This refers to the initial solution treatment and heat preservation time.

[0010] Optionally, the extraction of the thermal softening factor is based on the Arrhenius recovery kinetic equation, the expression of which is:

[0011] In the formula, , Let t be the absolute annealing temperature of the i-th pass. i Let i be the heat preservation time for the i-th pass. The activation energy is given by A and n, which are material kinetic constants, and R is the ideal gas constant.

[0012] Optionally, the expression for the strain-recovery alternating recursive equation is: ·(1-X) i ) +

[0013] In the formula, For the i-th intervention strain, X is the effective cumulative strain after the i-th pass. i The thermal softening factor for the i-th pass. This represents the effective cumulative strain after the (i-1)th pass.

[0014] Optionally, the method for outlier removal and standardization preprocessing of the original dataset includes: For all feature terms of the original dataset, calculate the sample mean μ and sample standard deviation σ of the single-dimensional data respectively; The normal data range is defined as [μ-3σ, μ+3σ]. Samples that exceed the normal data range are identified as outliers and are directly removed to obtain the effective dataset. For the effective dataset, normalize independently according to a single feature dimension, and convert all component parameters, multi-pass deformation parameters, annealing temperature and time parameters, mechanical properties, and electrical properties into standard data with a mean of 0 and a variance of 1.

[0015] Optionally, the machine learning prediction model uses extreme gradient boosting trees or random forests as the core algorithm, combined with L1 regularization constraints and decision tree depth limits, to reduce the risk of model overfitting for small sample high-noise data.

[0016] Optionally, the machine learning prediction model uses the coefficient of determination R² and root mean square error RMSE as quantitative evaluation indicators, and quantifies the contribution ratio of various mechanistic features to the prediction results through feature contribution analysis.

[0017] Optionally, the input features of the original dataset include the composition ratio of Cu-Ag alloy, the amount of cold deformation in each pass, the annealing temperature and the holding time, and the output features include electrical conductivity, hardness, yield strength, tensile strength and elongation at break.

[0018] Optionally, the expression for the determination coefficient R² is:

[0019] The expression for the root mean square error (RMSE) is:

[0020] Where m is the total number of samples in the test set. This represents the true performance value of the i-th sample in the experiment. The average of the true values ​​of all test samples. For the first i The model prediction performance value for each sample.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the industry pain points of existing Cu-Ag alloy performance prediction methods, which rely on pure data black-box models, lack support from microscopic evolution mechanisms, have poor interpretability, weak generalization ability for complex multi-pass processes, and involve high costs and long cycles in traditional experimental R&D. It innovatively constructs a dual-driven prediction system that couples multi-pass deformation and recovery evolution mechanisms with machine learning. For the first time, this invention introduces microscopic evolution mechanism features such as solid solution potential energy, thermal softening, strain accumulation, and dislocation strengthening into Cu-Ag alloy performance prediction modeling, completely breaking away from the traditional model's fitting mode that relies solely on macroscopic process parameters. This ensures that the model's prediction logic aligns with the metallurgical evolution laws of materials and possesses complete physical interpretability. Simultaneously, through data purification, regularized constraint modeling, and multi-dimensional model verification, it effectively solves the overfitting problem caused by small sample sizes and high noise in alloy experiments, significantly improving the model's prediction accuracy and generalization ability under unknown compositions and unknown complex multi-pass process paths. Attached Figure Description

[0022] Figure 1 This is a flowchart of the Cu-Ag alloy performance prediction method of the present invention; Figure 2 This is a fitting distribution diagram of the predicted values ​​and experimental real values ​​of the random forest model for the tensile strength of Cu-Ag alloy in this embodiment of the invention. The horizontal axis represents the experimental real values, and the vertical axis represents the model predicted values. Figure 3 This is a graph showing the fitting results of the XGBoost model for tensile strength prediction in an embodiment of the present invention; Figure 4 This is a graph showing the fitting effect of the conductivity random forest model in an embodiment of the present invention; Figure 5 The figure shows the fitting result of the conductivity XGBoost model in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Existing Cu-Ag alloy performance prediction technologies generally have significant drawbacks. Traditional experimental trial-and-error methods have long development cycles, high material costs, and extremely low efficiency in process optimization. Meanwhile, mainstream machine learning prediction schemes are mostly pure data black-box models that rely solely on fitting the original macroscopic process parameters without considering the microstructure evolution mechanism during multi-pass alloy processing. This results in a lack of physical interpretability of the models, insufficient generalization ability in complex multi-pass deformation and alternating annealing process scenarios, and prediction accuracy that is difficult to meet the needs of refined process control.

[0024] Based on this, the present invention breaks through the traditional single modeling mode, integrates material thermodynamics, kinetic mechanism and machine learning algorithm, and constructs a dual-input modeling system of original process dataset and microscopic mechanism feature dataset. Relying on physical mechanism to constrain the model fitting logic, it completely avoids the problem of pure data model having no theoretical support and being prone to spurious fitting.

[0025] See Figure 1 This invention provides a method for predicting the properties of Cu-Ag alloys by combining the characteristics of multi-stage deformation and recovery evolution, comprising the following steps: S1. Construct an original dataset covering the entire process of Cu-Ag alloy composition, process, and performance. Perform outlier removal and standardization preprocessing on the original dataset to obtain the first dataset.

[0026] Specifically, in the implementation of step S1, it is necessary to collect experimental data of Cu-Ag alloys with different composition ratios under various process conditions as the original dataset. The original dataset should cover composition parameters, including the content of major elements such as Cu and Ag, as well as other possible alloying elements; process parameters, including heat treatment and deformation process parameters such as cold deformation amount, annealing temperature, and holding time for each pass; and performance parameters, including mechanical and electrical performance indicators such as electrical conductivity, hardness, yield strength, tensile strength, and elongation at break.

[0027] It should be noted that the original dataset of this invention can be obtained in practical engineering applications either by consulting the materials science handbooks (1996-INFLUENCE OF SOLIDIFICATION CONDITIONS, THEMOMECHANICAL PROCESSING, AND ALLOYING ADDITIONS ON THE STRUCTURE AND PROPERTIES OF IN SITU COMPOSITE Cu-Ag ALLOYS) and empirical values ​​from literature, or by using historical experimental data from a small number of standard samples (such as Development of high-strength, high-conductivity Cu–Agalloys for high-field pulsed magnet use) for pre-fitting and calibration using mathematical tools such as nonlinear least squares; or by using the collected and standardized data for parameter optimization. Additionally, specific key parameters can also be obtained through a combination of a small number of experiments and algorithmic fitting.

[0028] The purpose of preprocessing is to remove noise and outliers from the data, thereby improving the stability and accuracy of subsequent model training. Specifically, this includes: Outlier removal: Identifying and removing data points that significantly deviate from the normal range using statistical methods.

[0029] Standardization preprocessing: Standardize features of different dimensions and orders of magnitude to eliminate the impact of numerical scale differences on model training.

[0030] In actual research and development, the final performance of Cu-Ag alloys is not determined by a single parameter. The proportion of silver, the amount of cold deformation processing each time, and the temperature and holding time of each annealing round all simultaneously change the internal structure of the material, ultimately affecting its electrical conductivity and mechanical strength. Therefore, it is necessary to first collect all experimental data corresponding to different composition formulations and different multi-pass processing routes to form a set of original datasets containing a complete correspondence between "composition-processing technology-finished product performance". The dataset records not only various processing input parameters, but also simultaneously includes the measured finished product performance data such as electrical conductivity and tensile strength.

[0031] During material testing, fluctuations in raw material purity, equipment temperature control errors, and deviations in sample testing operations can all generate outlier samples that deviate from the normal value range. Such outlier data can interfere with the subsequent model training convergence effect, so it is necessary to first perform outlier sample removal on the original dataset. At the same time, the numerical ranges and dimensions of various process parameters and performance indicators vary greatly. Directly inputting them into the model will cause an imbalance in feature weight allocation. After removing outlier samples, it is also necessary to perform standardization processing on all data to unify the numerical scale of data in each dimension. After processing, a standardized first dataset is obtained, which serves as the basic data carrier for subsequent modeling.

[0032] S2. Based on the heat treatment and multi-channel deformation and recovery softening feature data collected during the construction of the original dataset, calculations are performed to quantify and generate the mechanism features of the evolution of multiple micro-organisms, thus obtaining the second dataset.

[0033] Existing conventional machine learning prediction methods directly use macroscopic process parameters such as annealing temperature and cold deformation as model inputs, establishing only simple mapping relationships at the data level without relating to the actual microstructural evolution behavior that occurs during alloy processing. As a result, the interpretability of the model is poor, and the prediction error will increase significantly when a completely new alloy composition or a completely new processing route is used.

[0034] When Cu-Ag alloys undergo multi-pass rolling and alternating intermediate annealing, the matrix continuously undergoes microscopic changes such as solute solid solution, dislocation multiplication and accumulation, and recovery softening. This step does not directly use the original macroscopic process parameters, but instead combines the basic theoretical formulas of metal thermodynamics and kinetics, relying on the basic parameters related to solid solution, cold deformation, and annealing recorded in the original dataset, to calculate a series of characteristic parameters that can quantitatively characterize the solute solid solution state, the degree of annealing recovery softening, multi-pass cumulative strain, and dislocation strengthening contribution. These characteristic parameters all have clear material physical meanings and can quantify the mechanism characteristics of various microstructure evolutions. All mechanism characteristics are summarized to form a second dataset, which makes up for the deficiency of the original dataset, which only records macroscopic parameters and lacks microscopic evolution information.

[0035] S3. Based on the first and second datasets, construct and train machine learning prediction models for the target performance metrics respectively.

[0036] The Cu-Ag alloy exhibits various properties such as electrical conductivity, tensile strength, hardness, and elongation at break. Electrical conductivity primarily depends on the amount of silver solute atoms dissolved in the copper matrix, while tensile strength is dominated by the dislocation density accumulated from multiple deformation processes. Simultaneously predicting multiple properties using a single model is difficult to achieve accurate predictions for all indicators. Therefore, step S3 integrates the standardized original process characteristics with the converted microscopic evolution mechanism characteristics as model input. A dedicated model is built and trained for each predicted property, including electrical conductivity and tensile strength, allowing each model to adapt to the unique microscopic evolution laws of the corresponding property, thereby improving the prediction accuracy of individual performance indicators.

[0037] S4. The performance of the machine learning prediction model is evaluated by combining cross-validation and independent test sets, and the rationality and interpretability of the model mechanism are verified by feature analysis.

[0038] If the dataset is simply split into training and test sets for evaluation, the results will be subject to random bias. Therefore, step S4 uses a dual evaluation method of cross-validation and independent test sets. Cross-validation helps to avoid evaluation errors caused by random data splitting. In addition, an independent test set is constructed by selecting experimental samples that did not participate in the model training process to simulate real-world application scenarios with new alloy compositions and new processing techniques, objectively evaluating the model's prediction performance on unfamiliar samples.

[0039] S5. By inputting the Cu-Ag alloy composition ratio, cold deformation amount of each pass, annealing temperature and holding time of the Cu-Ag alloy to be tested into the machine learning prediction model, the output electrical conductivity, hardness, yield strength, tensile strength and elongation at break are obtained.

[0040] Using the machine learning prediction model trained in this application, staff only need to input the new alloy composition, the amount of deformation for each pass, the complete annealing temperature, and the holding time parameters to quickly obtain performance prediction data such as electrical conductivity, hardness, yield strength, tensile strength, and elongation at break, without the need for actual sample preparation or testing. Researchers can batch input multiple sets of alternative compositions and process parameters to obtain the corresponding performance prediction results at once, quickly selecting the process range that balances high strength and high electrical conductivity. This effectively reduces the number of physical tests, shortens the R&D cycle of high-strength, high-conductivity Cu-Ag alloys, and saves on raw materials, equipment, and labor costs during the R&D process.

[0041] In one feasible implementation, in step S2, calculations are performed based on the heat treatment and multi-pass deformation and recovery softening characteristic data collected during the construction of the original dataset to quantify and generate mechanistic features of multiple types of microstructure evolution, including: S201. Extract the initial solid solution potential energy characteristics.

[0042] The initial solution / homogenization treatment of Cu-Ag alloys determines the amount of silver atoms dissolved in the copper matrix, directly affecting the amount of subsequent precipitates, the electrical conductivity of the matrix, and the work hardening potential. Raw data only records the solution temperature and holding time, which cannot directly reflect the degree of solute atom diffusion saturation. Quantifying the solution potential energy using coupled thermodynamic kinetic formulas can characterize the initial solute distribution state in the matrix.

[0043] S202. Combine the single-pass annealing temperature, holding time and recovery activation energy to extract the thermal softening factor of each aging pass.

[0044] After each round of cold deformation, Cu-Ag alloys undergo annealing and aging. During annealing, recovery occurs, dislocations are annihilated, stored stress is released, and the work hardening effect of the previous pass is offset. The softening range varies greatly depending on the annealing temperature and holding time. The softening factor can quantify the degree of dislocation elimination by a single annealing pass, providing a softening correction coefficient for subsequent strain iteration calculations.

[0045] S203. By using the strain-recovery alternating recursive equation, the residual strain of the previous pass is corrected and the intervention strain of the current pass is superimposed.

[0046] The processing of Cu-Ag alloys exhibits strong path dependence, and the total deformation amount across all passes cannot be simply accumulated. Each annealing process eliminates some of the previously accumulated strain. By employing an alternating strain-recovery recursive equation, the true residual dislocation strain after alternating processing of "deformation-annealing softening-re-deformation" can be dynamically restored, accurately reflecting the deformation energy stored within the matrix.

[0047] S204. Nonlinear square root mapping is performed on the effective cumulative strain of multiple channels to obtain dislocation strengthening characteristics.

[0048] The core source of tensile strength enhancement in metallic materials is dislocation strengthening. The dislocation density has a square root nonlinear relationship with the effective cumulative strain. By taking the square root of the effective cumulative strain, the dislocation strengthening characteristics are obtained, which directly quantifies the strength contribution brought about by multiple processing passes, thus establishing a physical correlation between process parameters and mechanical properties.

[0049] In a feasible implementation, the method for extracting the initial solid solution potential energy characteristics in step S201 can be as follows: The initial solid solution potential energy characteristics are extracted using the Arrhenius equation, which couples thermodynamics and kinetics. The expression of the Arrhenius equation is as follows: (1) In equation (1), Let be the solid solution activation energy of the alloy, R be the ideal gas constant, k be the diffusion kinetic rate constant, and A1 be the material comprehensive pre-factor. The absolute temperature for solution / homogenization. This refers to the initial solution treatment and heat preservation time.

[0050] Specifically, this embodiment defines a quantitative calculation method for the initial solid solution potential energy characteristics and a parameter definition for the coupling equation, thus solving the technical defect that traditional techniques cannot quantitatively characterize the solid solution evolution law of Cu-Ag alloys.

[0051] The degree of solute solid solution saturation in Cu-Ag alloys directly determines the subsequent precipitation strengthening effect and the electrical conductivity of the matrix, and is the core fundamental factor affecting the final comprehensive performance of the alloy. Traditional single thermodynamic formulas can only characterize the temperature barrier effect and cannot reflect the dynamic process of solute diffusion driven by holding time, while single kinetic formulas ignore temperature energy constraints. Neither can accurately fit the coupling law of actual solid solution processes.

[0052] This invention employs the Arrhenius equation, which couples thermodynamics and kinetics, to simultaneously consider the energy barrier constraint of the solution temperature and the solute diffusion kinetics effect of the holding time. It can accurately quantify the solution potential energy level of the alloy under different solution temperatures and holding times, and achieve precise quantitative characterization of the initial matrix microstructure of the alloy.

[0053] In one feasible implementation, in step S202, the extraction of the thermal softening factor is based on the Arrhenius recovery kinetic equation, the expression of which is: (2) In equation (2), Let t be the absolute annealing temperature of the i-th pass. i Q is the heat preservation time for the i-th pass. rec To determine the activation energy, A and n are material kinetic constants, and R is the ideal gas constant, with a value of 8.314 J / (mol·K).

[0054] Specifically, this embodiment defines a kinetic calculation method for the thermal softening factor of multi-pass annealing, solving the technical blind spot of traditional process modeling being unable to quantify the softening differences of multi-pass intermittent annealing. The industrial preparation of Cu-Ag alloys generally adopts an alternating process of multi-pass deformation and intermittent annealing. The temperature and holding time of each annealing pass are different, which directly leads to different degrees of softening in dislocation annihilation, stress release, and microstructure recovery.

[0055] Traditional prediction models only use statistical modeling of overall process parameters to qualitatively describe the softening effect of annealing, failing to distinguish the microscopic softening differences between each annealing pass, which does not match the actual metallurgical evolution process. This embodiment, based on the Arrhenius recovery kinetic equation and combined with the annealing process parameters and alloy recovery activation energy of each pass, quantitatively calculates the thermally induced softening factor for each pass, accurately characterizing the microstructure recovery and strain softening effect corresponding to each round of annealing, and achieving a refined and differentiated quantitative characterization of the behavior of multi-pass intermittent heat treatment.

[0056] In one feasible implementation, in step S203, the expression for the strain-recovery alternating recursive equation is: ·(1-X) i ) + (3) In equation (3), For the i-th intervention strain, X is the effective cumulative strain after the i-th pass. i The thermal softening factor for the i-th pass. This represents the effective cumulative strain after the (i-1)th pass.

[0057] Traditional alloy performance prediction models often use total cold deformation as the strain input parameter, ignoring the alternating cycle of "strain hardening and annealing softening" in multi-pass processing. This fails to reflect the inherited characteristics of the process path, resulting in severe distortion of strain accumulation characterization, which does not match the true microscopic state of the material.

[0058] This invention employs a recursive iterative calculation logic. Based on the residual effective strain of the previous pass, after the annealing and softening attenuation of the current pass, the cold deformation intervention strain of the current pass is superimposed. This accurately replicates the dynamic evolution law of strain accumulation and recovery attenuation during multi-pass processing, and obtains the effective accumulated strain that conforms to the true microscopic state of the material.

[0059] Suppose a material undergoes three processing steps: First round: =0.5, no annealing, X1=0, = 0 + 0.5 = 0.5; Second round: =0.3, Annealing softening x2 = 0.2, = 0.5×(1-0.2)+0.3 =0.4+0.3=0.7; Third round: =0.4, Annealing and softening x3 = 0.3, = 0.7×(1-0.3)+0.4= 0.49+0.4=0.89.

[0060] This recursive calculation allows for accurate tracking of the strain evolution history of materials during multiple processing steps.

[0061] In one feasible implementation, step S1, the method for outlier removal and standardization preprocessing of the original dataset, includes: S101. For all feature items of the original dataset, calculate the sample mean μ and sample standard deviation σ of the single-dimensional data respectively.

[0062] Experimental data on Cu-Ag alloys are easily affected by experimental conditions, equipment errors, and human error, resulting in significant differences in data dispersion across various characteristic dimensions. Without precise single-dimensional statistical parameters, there is no quantitative basis for subsequent outlier identification. This step uses standardized mathematical statistics to provide a unified and objective quantitative benchmark for defining outlier intervals and identifying data biases, ensuring that subsequent data cleaning work has a rigorous mathematical basis and avoiding dataset distortion caused by subjective selection.

[0063] S102. Set the normal data range as [μ-3σ, μ+3σ]. Identify samples that exceed the normal data range as outliers and remove them directly to obtain the valid dataset.

[0064] Cu-Ag alloy experiments are susceptible to extreme anomalies caused by factors such as raw material purity, rolling equipment, temperature control errors, and sample testing deviations. A single anomaly can severely interfere with model weight learning. By calculating the mean and standard deviation for each feature, only samples within three standard deviations are retained, and entire anomalies are directly removed, thus reducing dataset noise at its source.

[0065] S103. For the valid dataset, normalize independently according to a single feature dimension, and convert all component parameters, multi-pass deformation parameters, annealing temperature and time parameters, mechanical properties, and electrical properties into standard data with a mean of 0 and a variance of 1.

[0066] The original features exhibit vastly different numerical magnitudes: silver content ranges from a few tenths of a percent, deformation from 0% to 90%, annealing temperature from several hundred K, tensile strength from 1000 MPa to 1500 MPa, and electrical conductivity from 60 IACS to 100 IACS. These different numerical dimensions can cause machine learning models to overemphasize the weights of large numerical parameters and neglect key small numerical parameters. Standardizing each dimension to a mean of 0 and a variance of 1 can eliminate training bias caused by these different dimensions.

[0067] In a feasible implementation, the aforementioned machine learning prediction model can use Extreme Gradient Boosting (XGBoost) or Random Forest (RF) as the core algorithm, and can be combined with L1 regularization constraints and decision tree depth limits to reduce the risk of model overfitting on small sample high-noise data.

[0068] The experimental research and development of Cu-Ag alloys faces high costs and limited sample sizes, with significant nonlinear coupling characteristics in the data. Traditional single machine learning algorithms exhibit poor stability, are prone to overfitting and spurious fitting, and have weak generalization capabilities, making them unsuitable for predicting unknown operating conditions. This embodiment utilizes XGBoost and Random Forest ensemble learning algorithms, which possess strong nonlinear fitting capabilities and excellent anti-interference performance, to adapt to complex input systems with coupled multi-mechanism features. Simultaneously, L1 regularization constraints and decision tree depth limitations are applied to proactively suppress data noise interference at the model structure and parameter levels, effectively mitigating the risk of overfitting in small-sample modeling.

[0069] In a feasible implementation, the machine learning prediction model can use the coefficient of determination R² and the root mean square error RMSE as quantitative evaluation indicators. The contribution ratio of various mechanistic features to the prediction results can be quantified through feature contribution analysis.

[0070] Traditional techniques rely solely on a single accuracy metric to evaluate model performance, failing to comprehensively verify the model's ability to fit patterns and its error stability. Furthermore, the model's predictive logic is closed, and feature contributions are untraceable, lacking theoretical persuasiveness and making it difficult to demonstrate technological innovation. This embodiment employs a dual-metric evaluation using the R² coefficient of determination and the root mean square error (RMSE) to comprehensively verify model performance from two dimensions: the fit to the performance evolution pattern and the accuracy of absolute prediction error. This achieves a standardized and comprehensive evaluation of model accuracy. Simultaneously, a feature contribution analysis method is introduced to quantitatively quantify the predictive contribution ratio of self-developed mechanism features such as solid solution potential energy, thermal softening factor, and effective cumulative strain.

[0071] In one feasible implementation, the original dataset in step S1 can be input with features including Cu-Ag alloy composition ratio, cold deformation amount of each pass, annealing temperature and holding time, and can output features including electrical conductivity, hardness, yield strength, tensile strength and elongation at break.

[0072] Most existing alloy prediction technologies only model single performance indicators and single process parameters, with a narrow feature coverage, which cannot meet the diversified R&D needs of high-strength and high-conductivity Cu-Ag alloys and have limited engineering application value.

[0073] This embodiment fully covers four major categories of core process input characteristics: component ratio, multi-pass cold deformation amount, annealing temperature, and holding time. It also covers mechanical and electrical multi-dimensional output performance indicators such as electrical conductivity, hardness, yield strength, tensile strength, and elongation at break. It meets the actual needs of industrial R&D, preparation, and performance testing of Cu-Ag alloys and can achieve multi-parameter collaborative input and simultaneous and accurate prediction of multi-target performance.

[0074] In one feasible implementation, the determination coefficient R² is expressed as follows: (4) The expression for the root mean square error (RMSE) is: (5) In equations (4) and (5), m is the total number of samples in the test set. This represents the true performance value of the i-th sample in the experiment. The average of the true values ​​of all test samples. For the first i The model prediction performance value for each sample.

[0075] Specifically, this embodiment defines the standard calculation formulas for the core model evaluation metrics R² and RMSE, unifying the quantitative standards for model performance evaluation and making the model validation system standardized, reproducible, and verifiable. Without clear formula definitions, model evaluation standards would be inconsistent, and comparisons of prediction accuracy would lack rigorous mathematical basis, making it impossible to intuitively and objectively demonstrate the technical advantages of this invention. This embodiment, by adopting standardized calculation formulas, can more clearly define the physical definitions and calculation logic of each parameter, thereby achieving the standardization and normalization of model evaluation.

[0076] It should be noted that in step S204, the... Nonlinear mapping leads to the final dislocation enhancement features. The expression is: (6) The growth rate of dislocation strengthening in metals gradually slows down with increasing strain, and is not purely linearly proportional to cumulative strain. Directly modeling with effective cumulative strain would violate the true physical laws governing deformation strengthening in Cu-Ag alloys.

[0077] This invention uses square root nonlinear mapping to transform the effective cumulative strain at the macroscopic level into a quantitative characteristic of the microscopic dislocation strengthening level. Through this embodiment, the alloy deformation law of "rapid growth of strengthening at small strain and tendency to saturation of strengthening at large strain" can be fitted.

[0078] Before constructing the original dataset covering the entire process of Cu-Ag alloy composition, process, and performance in step S1, it also includes: comprehensively collecting experimental data of Cu-Ag alloys under different composition ratios and complex process paths.

[0079] This step involves collecting a wide range of experimental samples covering diverse process paths, including different Ag content ratios, different solution treatment regimes, multiple cold deformation passes, and multiple annealing temperatures and holding times. This ensures that the typical process windows and compositional ranges in the actual preparation of Cu-Ag alloys are fully covered, avoiding the problem of insufficient model generalization ability due to single data and incomplete coverage of working conditions.

[0080] The "training the machine learning prediction model" mentioned in step S3 involves the following steps: The merged dataset (combining the first and second datasets) is divided into a training set, a validation set, and an independent test set in a 7:2:1 ratio. RF and XGBoost are selected as the basic algorithms, and overfitting is suppressed by setting the number of decision trees, maximum tree depth, learning rate, and L1 regularization constraints. Taking tensile strength prediction as an example, the model input includes alloy composition ratios, multi-pass process parameters, standardized data, and microscopic mechanism features such as initial solid solution potential, thermal softening factor, effective cumulative strain, and dislocation strengthening characteristics. The output includes tensile strength, yield strength, hardness, electrical conductivity, and elongation. Iterative learning is performed using the training set, and parameter optimization is performed using the validation set until the model converges and there is no overfitting, ultimately yielding the trained performance prediction model.

[0081] To further illustrate the prediction method of the present invention, the present invention will be described in conjunction with specific embodiments.

[0082] Example 1 This embodiment takes the prediction of electrical conductivity (IACS%) and tensile strength (MPa) of Cu-Ag alloy as an example to illustrate the entire process from model establishment to performance prediction.

[0083] In this embodiment, following the Cu-Ag alloy performance prediction method described above, RF and XGBoost machine learning models were established for the Cu-Ag alloy electrical conductivity (IACS%) and tensile strength Rm (MPa).

[0084] Before constructing the original dataset, we comprehensively collected experimental data of Cu-Ag alloys under different composition ratios and complex process paths. The samples cover multiple gradient Ag composition ratios, multiple cold deformation regimes, different combinations of annealing temperatures and holding times, and cover typical process windows in actual industrial production, ensuring that the samples are comprehensive, representative and process-covering.

[0085] Subsequently, an original dataset covering the entire process of Cu-Ag alloy composition, process, and performance was constructed, and outlier removal and standardization preprocessing were performed according to the above method to obtain the first dataset. Then, according to the above mechanism feature calculation method, the initial solid solution potential energy feature extraction, thermal softening factor calculation for each pass, strain-recovery alternating recursive solution of effective cumulative strain, and dislocation strengthening feature obtained by nonlinear square root mapping were completed to construct the second dataset.

[0086] Furthermore, based on the first and second datasets, machine learning prediction models were constructed and trained for the target performance metrics. The specific training method for the models is as follows: The first and second datasets were merged and randomly divided into training, validation, and independent test sets in a 7:2:1 ratio. XGBoost and RF were selected as the core algorithms, along with L1 regularization constraints and decision depth limits to reduce the risk of overfitting caused by small sample sizes and high noise. The model input includes both the original alloy processing characteristics and microscopic mechanism characteristics such as solid solution potential energy, thermal softening factor, effective cumulative strain, and dislocation strengthening. A dedicated prediction model was independently constructed for each performance indicator—electrical conductivity, tensile strength, yield strength, hardness, and fracture elongation—to avoid feature interference caused by mixed training of multiple indicators.

[0087] Taking the training of the tensile strength prediction model as an example, the model input includes the composition ratio of the Cu-Ag alloy, the cold deformation amount of each pass, the annealing process parameters, as well as the initial solid solution potential energy, thermal softening factor, effective cumulative strain, and dislocation strengthening characteristics; the model output is the measured value of the alloy's tensile strength. During training, the training set is used to iteratively learn the nonlinear correlation between composition, process, microstructure, and tensile strength. The model fitting status is monitored in real time through the validation set. Training is terminated when the accuracy of the validation set tends to stabilize and no longer improves, and the optimal model parameters are retained. The electrical conductivity and other performance models adopt the same training process, only changing the corresponding performance output labels to obtain the specific prediction model for each target performance.

[0088] After model training, a combination of cross-validation and independent test sets was used to evaluate performance using R² and RMSE as evaluation metrics. For the tensile strength XGBoost and RF models, R² reached 0.9470 and 0.9533, respectively, with RMSEs of 66.3286 and 62.2262. For the conductivity XGBoost and RF models, R² reached 0.9008 and 0.8938, respectively, with RMSEs of 2.9378 and 3.0403. Compared to traditional models that only use the original process parameters, the optimized model of this invention, which incorporates microscopic evolution mechanism features, shows significant improvements in both goodness of fit and prediction accuracy.

[0089] Figures 2-5The results of training and prediction for different machine learning models are presented.

[0090] Figure 2 This is a plot showing the fitted distribution of predicted and experimental values ​​for the random forest model of tensile strength in Cu-Ag alloys, based on the training and test sets. The horizontal axis represents the experimental values, the vertical axis represents the model predictions, and the diagonal line represents the ideal fitting baseline. Figure 2 As can be seen, both the training and test samples are closely distributed on both sides of the baseline, with no obvious discrete deviations, indicating that the RF model has a good fitting effect on tensile strength, with high overall correlation and good fitting stability.

[0091] Figure 3 The results show the fitting of the XGBoost model for tensile strength prediction. The samples are uniformly distributed along the ideal fitting baseline with low dispersion. Compared to traditional models, the XGBoost model introduced with mechanistic features in this invention has good fitting ability for samples in different strength ranges, with no obvious overfitting or underfitting, and excellent generalization performance.

[0092] Figure 4 The figure shows the fitting effect of the conductivity RF model. The sample points are concentrated and closely follow the baseline distribution. The high and low conductivity ranges maintain good fit consistency, indicating that the model can accurately characterize the nonlinear relationship between composition, process and conductivity, with small prediction deviation and high fit.

[0093] Figure 5 The results of fitting the XGBoost model for conductivity show that the sample distribution is compact with few deviations. The model can achieve high-precision prediction across the entire range, demonstrating that the mechanism + data dual-driven feature constructed in this invention can effectively characterize the microscopic evolution of the conductivity of Cu-Ag alloys.

[0094] Figures 2-5 Both the training and test samples closely approximate the ideal fitting baseline, with few discrete outliers, proving that both RF and XGBoost algorithms can achieve high-precision fitting based on the mechanistic feature dataset of this invention. Tensile strength and electrical conductivity show good fitting across the entire numerical range without interval bias, indicating that the mechanistic features extracted by this invention, such as solid solution, strain, and dislocation, can truly reflect the intrinsic correlation between the process, microstructure, and macroscopic properties of Cu-Ag alloys. This invention effectively improves the model's fitting accuracy and generalization ability through multi-pass deformation and recovery evolution feature modeling, possessing reliable engineering prediction and process optimization value.

[0095] Compared to calculations without incorporating multi-pass deformation and recovery evolution characteristics... and the final dislocation enhancement features The optimization performance of the models trained on the datasets is shown in Table 1: Table 1 Comparison of prediction accuracy of tensile strength Rm model before and after optimization

[0096] Similarly, for the conductivity prediction model, the R² coefficients of determination for RF and XGBoost are 0.8938 and 0.9008, respectively, and the RMSEs are 3.0403 and 2.9378, respectively. The comparison of prediction accuracy before and after optimization is shown in Table 2. Table 2 Comparison of prediction accuracy before and after conductivity model optimization

[0097] Specific examples of model prediction validation are compared in detail in Comparative Example 1.

[0098] Comparative Example 1 To verify the practical technical effect of the method of this invention, the performance prediction of Cu-Ag alloys was performed using both the traditional RF model and the optimized RF model of this invention. Independent test set data that was not used in the training were randomly selected, and their feature matrices were input into the traditional RF model and the optimized RF model, respectively, to obtain the performance prediction output. Subsequently, the prediction results of the two models were cross-compared with the actual experimental test values ​​to quantitatively verify the significant advantages of the optimized scheme of this invention in reducing prediction errors and improving generalization ability. The specific results are shown in Table 3.

[0099] Table 3. Predicted tensile strength Rm using the traditional RF model.

[0100] Table 4. Prediction results of tensile strength Rm of the optimized RF model

[0101] Table 5. Predicted conductivity results of the traditional RF model

[0102] Table 6. Predicted conductivity results of the optimized RF model

[0103] By comparing the prediction results of the traditional RF model and the optimized RF model in Tables 3 to 6 regarding tensile strength Rm and electrical conductivity, it can be seen that: 1. In terms of tensile strength prediction, traditional RF models have many problems with excessive prediction bias, with relative errors of over 20% for some samples, large dispersion and poor generalization ability; the optimized RF model of this invention, after introducing the characteristics of multi-channel deformation and recovery evolution mechanism, has a significantly reduced overall absolute error, with the relative error of most samples controlled within 10%, and the prediction stability is greatly improved.

[0104] 2. In terms of conductivity prediction, the traditional RF model also has the defects of high error in some samples and insufficient consistency of fitting; the optimized RF model has a higher degree of fit between the predicted value and the actual experimental value, with the relative error of most samples controlled within 5%, and the proportion of low error samples significantly increased.

[0105] 3. Overall comparison shows that this invention, by introducing microscopic mechanism features such as initial solid solution potential energy, thermal softening factor, effective cumulative strain, and dislocation strengthening, constructs a dual-driven prediction system based on mechanism and data. This effectively compensates for the shortcomings of traditional models that rely solely on macroscopic process parameters and lack constraints from material evolution mechanisms. It significantly reduces the absolute and relative errors in prediction, effectively improves the model's fitting accuracy and generalization ability, and can more accurately and reliably predict the mechanical and electrical properties of Cu-Ag alloys under different compositions and process paths. It has obvious technical advantages and practical engineering value.

[0106] The above description is merely the preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the properties of Cu-Ag alloys by combining the characteristics of multi-stage deformation and recovery evolution, characterized in that, Includes the following steps: A raw dataset covering the entire process of Cu-Ag alloy composition, process, and performance is constructed. Outlier removal and standardization preprocessing are performed on the raw dataset to obtain the first dataset. Based on the heat treatment and multi-pass deformation and recovery softening feature data collected during the construction of the original dataset, calculations are performed to quantify and generate the mechanistic features of multiple types of microstructure evolution, resulting in a second dataset. Based on the first dataset and the second dataset, machine learning prediction models are constructed and trained for the target performance metrics, respectively. The performance of the machine learning prediction model was evaluated by combining cross-validation and independent test sets, and the rationality and interpretability of the model mechanism were verified by feature analysis. By inputting the Cu-Ag alloy composition ratio, cold deformation amount of each pass, annealing temperature and holding time of the Cu-Ag alloy to be tested into the machine learning prediction model, the electrical conductivity, hardness, yield strength, tensile strength and elongation at break are output.

2. The method for predicting the properties of Cu-Ag alloys according to claim 1, characterized in that, The heat treatment and multi-pass deformation and recovery softening feature data collected during the construction of the original dataset are used to calculate and quantify the mechanistic features of multiple types of microstructure evolution, including: Initial solid solution potential energy characteristics were extracted. By combining the single-pass annealing temperature, holding time, and recovery activation energy, the thermo-softening factor of each aging pass was extracted. The strain-recovery alternating recursive equation is used to correct the residual strain of the previous pass and superimpose the intervention strain of the current pass. Nonlinear square root mapping of effective cumulative strain across multiple channels yields dislocation strengthening characteristics.

3. The method for predicting the properties of Cu-Ag alloys according to claim 2, characterized in that, The method for extracting the initial solid solution potential energy characteristics is as follows: The initial solid solution potential energy characteristics are extracted using the Arrhenius equation, which couples thermodynamics and kinetics. The expression of the Arrhenius equation is as follows: In the formula, Let be the solid solution activation energy of the alloy, R be the ideal gas constant, k be the diffusion kinetic rate constant, and A1 be the material comprehensive pre-factor. The absolute temperature for solution / homogenization. This refers to the initial solution treatment and heat preservation time.

4. The method for predicting the properties of Cu-Ag alloys according to claim 2, characterized in that, The extraction of the thermal softening factor is based on the Arrhenius recovery kinetic equation, the expression of which is: In the formula, , Let t be the absolute annealing temperature of the i-th pass. i The heat preservation time for the i-th pass is... The activation energy is given by A and n, which are material kinetic constants, and R is the ideal gas constant.

5. The method for predicting the properties of Cu-Ag alloys according to claim 2, characterized in that, The expression for the strain-recovery alternating recursive equation is: ·（1-X i ） + In the formula, For the i-th intervention strain, X is the effective cumulative strain after the i-th pass. i The thermal softening factor for the i-th pass. This represents the effective cumulative strain after the (i-1)th pass.

6. The method for predicting the properties of Cu-Ag alloys according to any one of claims 2-5, characterized in that, The method for outlier removal and standardization preprocessing of the original dataset includes: For all feature terms of the original dataset, calculate the sample mean μ and sample standard deviation σ of the single-dimensional data respectively; The normal data range is defined as [μ-3σ, μ+3σ]. Samples that exceed the normal data range are identified as outliers and are directly removed to obtain the effective dataset. For the effective dataset, normalize independently according to a single feature dimension, and convert all component parameters, multi-pass deformation parameters, annealing temperature and time parameters, mechanical properties, and electrical properties into standard data with a mean of 0 and a variance of 1.

7. The method for predicting the properties of Cu-Ag alloys according to claim 6, characterized in that, The machine learning prediction model uses Extreme Gradient Boosting Tree (XGBoost) or Random Forest (RF) as the core algorithm, combined with L1 regularization constraints and decision tree depth limits to reduce the risk of model overfitting on small sample high-noise data.

8. The method for predicting the properties of Cu-Ag alloys according to claim 6, characterized in that, The machine learning prediction model uses the coefficient of determination R² and root mean square error RMSE as quantitative evaluation indicators. Through feature contribution analysis, it quantifies the contribution ratio of various mechanistic features to the prediction results.

9. The method for predicting the properties of Cu-Ag alloys according to claim 6, characterized in that, The input features of the original dataset include the composition ratio of Cu-Ag alloy, the amount of cold deformation in each pass, the annealing temperature and the holding time, and the output features include electrical conductivity, hardness, yield strength, tensile strength and elongation at break.

10. The method for predicting the properties of Cu-Ag alloys according to claim 8, characterized in that, The expression for the determination coefficient R² is: The expression for the root mean square error (RMSE) is: Where m is the total number of samples in the test set. This represents the true performance value of the i-th sample in the experiment. The average of the true values ​​of all test samples. For the first i The model prediction performance value for each sample.

Images